I summarized the paper at the following URL in my own way.
「PrivacyMic: Utilizing Inaudible Frequencies for Privacy Preserving Daily Activity Recognition」
https://dl.acm.org/doi/pdf/10.1145/3411764.3445169
CTAC 2024 Valencia - Sven Zoelle - Most Crucial Invest to Digitalisation_slid...
Summary of the paper「PrivacyMic: Utilizing Inaudible Frequencies for Privacy Preserving Daily Activity Recognition」
1. PrivacyMic: Utilizing Inaudible
Frequencies for Privacy Preserving Daily
Activity Recognition
Iravantchi, Yasha and Ahuja, Karan and Goel, Mayank and
Harrison, Chris and Sample, Alanson
Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems
2021/06/23
Yuta Funada (M1)
2. Overview
• In order to improve the privacy of sound-based sensing
devices, we proposed a microphone that recognizes human
behavior using sounds outside the frequency range of human
speech.
• We conducted a perception study, where participants
“eavesdropped” on PrivacyMicʼs filtered audio and found that
none of our participants could transcribe speech.
• Over 95% classification accuracy in any environment.
3. RELATED WORK
• Collected Sound Databases and Labeling
• Acoustic Activity Recognition
• Infrasound and Ultrasound for Sensing
• Privacy and Microphones
No previous studies using ultrasound.
4. INAUDIBLE SOUNDS COLLECTION
• Do our daily-use objects emit significant infrasonic and
ultrasonic sounds?
• If the devices do emit these sounds, are these inaudible
frequencies useful for recognition?
5. Wideband Capture Apparatus
• To collect sounds from three distinct
regions of the acoustic spectrum, we
built an audio-capture rig that combines
three microphones with targeted frequency
responses: infrasound, audible, and ultrasound.
6. Data Collection Procedure
• To introduce real-world variety and many different objects,
including different models of the same item (e.g., Shark
vacuum vs. Dyson vacuum), data was collected across three
homes and four commercial buildings.
• Before recording the object, a 5-second snapshot was taken
as a background recording to be used later for background
subtraction.
9. Featurization
• Create a high-resolution FFT
• Subtracting the background FFT components from the objectʼs FFT
• Composite a 100 log-binned feature vector from 0Hz to 192kHz.
10. Spectral Information Power
• Employ feature selection methods that rank each band by its
information power
• Using Random Forests
11. Classification Accuracy
• We quantify our results of spectral analysis in terms of
classification accuracies as well
• Our evaluation, we use a Random Forest Classifier with 1000
estimators
12. HARDWARE IMPLEMENTATION
• Set out to design a microphone optimized for high-audible
and ultrasonic frequencies
• PrivacyMic consists of three major components:
・a wide-band ultrasonic microphone
・in-hardware amplification and filter stage
・a low-noise, low power high speed Analog to Digital Converter (ADC)
14. Hardware Evaluation
• Filter Performance
• Distance Performance
- Even at a distance of 15 meters, we were
able to pick up more signal than noise.
• Power Consumption
- sustain continuous operation for 7 hours.
(1600mAh battery)
• Wireless Performance
- an average overhead of 2ms on top of
the network latency.
Filter Performance
Distance Performance
⾳声フィルタ(8k)と可聴域フィルター(16k)
15. PRIVACY EVALUATION
• There are numerous privacy concerns microphones in our homes
• Privacy risk situations
- bad actors gain access to audio streams of the device directly
- through mishandled data breaches
• Evaluating whether our participants were able to perceive
various levels of content within a series of audio clips
16. Procedure
• We generated 3 audio files by reading a selected passage from Wikipedia
• For file A, we used our speech filter, removing all frequencies below 8kHz
• For file B, we pitch shifted the harmonic frequencies down to 300Hz
• For file C, we used our audible filter, removing all frequencies below
16kHz
17.
18. Results
• For file A(8kHzまでカット)
- Could not hear human sounds
- None were able to transcribe a single word
• For file B(Aで残った⾳声の断⽚をピッチシフト)
- Could hear something, and like a human sounds
- None were able to transcribe a single word
• For file C(16kHzまでカット)
- A fewer number of participants answered that they could
hear sounds
- None were able to transcribe a single word
19. REAL WORLD PERFORMANCE
• We also evaluated the performance of PrivacyMic in a less
controlled environment.
21. DISCUSSION
• In many systems, there is a trade-off between accuracy and
privacy. However, because there was a wealth of information
present in the ultrasound, even with the 20 kHz cutoff, the
loss of accuracy was only 5%.
• If you are "listening" to audio recorded by PrivacyMic, you
will find that humans cannot understand speech.
• We expect future eavesdropping attacks to more likely be
algorithmic.
• Increase the cutoff value of the filter.
23. RELATED WORK
• Collected Sound Databases and Labeling
• Acoustic Activity Recognition
• Infrasound and Ultrasound for Sensing
• Privacy and Microphones